Cross-domain federated machine learning training method oriented to image omics

Abstract

The invention relates to the technical field of federated learning, and discloses an image omics-oriented cross-domain federated machine learning training method, which comprises the following steps that: a server distributes a global orthogonal semantic base to establish a shared gradient space recessive semantic coordinate system, and when the stability of a local training process meets a preset condition, a client sends the global orthogonal semantic base to the client; a local gradient is projected to a coordinate system and uploaded together with a digital fingerprint representing the physical characteristics of imaging equipment, and a server groups clients according to the digital fingerprint and performs weighted aggregation based on the information concentration degree on received projection vectors in each group so as to reconstruct a global gradient to update a model. According to the method, through an aggregation mode of firstly performing grouping according to physical causes and then performing weighting in groups according to information values, systematic deviation from equipment differences and real pathological semantics from data contents can be effectively stripped, and sensitivity to key clinical features is kept.

Description

Technical Field

[0001] This invention relates to a cross-domain federated machine learning training method for image omics, belonging to the field of federated learning technology. Background Technology

[0002] Currently, in order to utilize heterogeneous data from multiple centers while protecting data privacy, a mainstream technical approach has emerged in the industry: logically constructing a unified view of data through secure multi-party computation or feature space alignment.

[0003] However, when this technology is applied to large-scale and cross-regional real-world medical collaboration scenarios, such as building a nationwide disease screening network covering thousands of hospitals, the hidden costs of maintaining its operation become apparent. On the one hand, the complex cryptographic operations used to ensure privacy and security place high demands on the computing power of some nodes in the network, increasing deployment costs. On the other hand, in order to cope with the data heterogeneity caused by different brands and models of imaging equipment, forcibly aligning and fusing at the feature level may result in the loss of some information reflecting the specificity of regional cases during the fusion process.

[0004] To address the aforementioned issues, the industry has attempted to simplify privacy-preserving algorithms or develop more complex feature normalization models. However, the former sacrifices security, while the latter further increases the computational burden and implementation complexity of the system, failing to fundamentally resolve the inherent contradictions of existing methods. Specifically, existing technologies suffer from the following shortcomings: 1. There is an irreconcilable conflict of computational resources between ensuring high-level privacy and security and achieving low-cost participation of large-scale nodes; 2. There is an inherent contradiction of information loss between pursuing global feature space consistency and protecting the unique diagnostic value of local data of each node; 3. There is a lack of a unified framework that can simultaneously address the three heterogeneous challenges of systemic biases arising from device physical differences, statistical differences arising from data content distribution, and temporal feature drift arising from disease evolution. Therefore, the technical problem to be solved by this invention is how to design a federated learning training method that can achieve implicit semantic alignment of cross-domain heterogeneous model updates without relying on high-cost explicit alignment and complex encryption, and how to make the method inherently robust to multi-dimensional challenges such as differences in device physical characteristics, uneven distribution of client data, and dynamic drift of local data. Summary of the Invention

[0005] This invention provides a cross-domain federated machine learning training method for radiomics. Its main purpose is to solve the problems that existing federated learning methods are limited in practicality when realizing cross-domain collaboration due to their reliance on high-cost explicit alignment and complex encryption, and lack a unified solution mechanism to address the multi-dimensional heterogeneous challenges such as physical differences in devices, uneven data distribution, and dynamic feature drift.

[0006] To achieve the above objectives, this invention provides a cross-domain federated machine learning training method for image omics. The method is applied to a federated learning system comprising a server and multiple clients, and includes:

[0007] The server generates a set of globally orthogonal semantic basis vectors and distributes them to multiple clients to establish a shared gradient space implicit semantic coordinate system.

[0008] Each client performs the following steps: Before training the local model, a standardized digital micro-perturbation signal is superimposed on a digital copy of the local medical image data to generate a perturbation image, and the gradients of the feature maps generated by the local medical image data and the perturbation image are extracted respectively, and then the difference in gradient response between the two is calculated to obtain a digital fingerprint.

[0009] Models are trained based on their respective local data to obtain local gradient vectors. During local model training, a stability metric representing the curvature change of the loss function is obtained, and it is determined whether the stability metric is within a preset stable range. If the result is yes, the local gradient vector is projected onto each basis vector of the global orthogonal semantic basis to obtain a low-dimensional semantic projection vector. The digital fingerprint and the low-dimensional semantic projection vector are uploaded to the server, which performs the following steps: performing clustering operations on the digital fingerprints received from multiple clients to divide the clients into one or more device clusters; within each device cluster, weighted aggregation of the low-dimensional semantic projection vectors uploaded by the clients within that device cluster is performed to generate the intra-cluster average semantic projection vector of each device cluster, where each weight is calculated based on the concentration metric corresponding to the component distribution of the low-dimensional semantic projection vector. Finally, the server fuses the intra-cluster average semantic projection vectors and combines them with the global orthogonal semantic basis to reconstruct the global average gradient, which is used to update the global model.

[0010] Preferably, the concentration measure is information entropy. Weights and Information Entropy Inversely proportional, information entropy Determined by the following formula: ,in, The number of globally orthogonal semantic basis vectors. For the client In the low-dimensional semantic projection vector, the first The probability value is obtained by normalizing the absolute value of each component.

[0011] Preferably, the standardized digital micro-perturbation signal is a single-pixel point pulse signal; the digital fingerprint is the vector magnitude of the gradient response difference.

[0012] Preferably, the stability metric is the variance of the loss function curvature within a sliding time window; the preset stability interval is when the variance is less than a preset drift trigger threshold.

[0013] Preferably, when the stability metric exceeds a preset stability range, the client pauses uploading low-dimensional semantic projection vectors and sends a resynchronization request to the server to obtain the latest global model.

[0014] Preferably, the clustering operation performed by the server is the K-Means clustering algorithm.

[0015] Preferably, the step of the client projecting its local gradient vector onto each basis vector of the global orthogonal semantic basis is as follows: calculate the dot product of the local gradient vector and each basis vector to obtain the corresponding projection length as a component of the low-dimensional semantic projection vector.

[0016] Preferably, the step of server fusion of average semantic projection vectors within each cluster includes: performing a weighted average of the average semantic projection vectors within each cluster to generate a fused average semantic projection vector, wherein the weights used for weighting the average are proportional to the number of clients contained in the cluster.

[0017] Preferably, the global orthogonal semantic basis vectors are generated by the server during the initialization phase of the federated learning system by performing a Gram-Schmidt orthogonalization process on a set of randomly generated vectors.

[0018] Preferably, the step of the server reconstructing the global average gradient is as follows: multiply each component of the fused average semantic projection vector by the basis vector of its corresponding global orthogonal semantic basis, and then sum all the resulting vectors.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. By applying standardized digital micro-perturbation signals locally on the client and analyzing their gradient response differences, a digital fingerprint characterizing the physical characteristics of each client's imaging device is obtained. This allows the server to effectively cluster clients with similar imaging physical characteristics based on this digital fingerprint before performing aggregation operations. Then, within each device cluster, the weight in the weighted aggregation is determined based on the concentration measure of the component distribution of each low-dimensional semantic projection vector. This two-tiered aggregation method, which first groups based on physical causes and then weights them within the cluster based on information value, enables this method to effectively distinguish between systematic biases arising from differences in device physical characteristics and true semantic information arising from the distribution of data content. It avoids feature confusion caused by mixing gradient information from different physical sources. At the same time, it ensures that the contributions of specialized nodes carrying rare case information are given due attention in homogeneous device clusters. Thus, the final global model maintains both generalization adaptability to different hardware environments and sensitivity to key clinical features.

[0021] 2. During the execution of the method of this invention, each client can continuously monitor the change in the curvature of the loss function during the training of its local model, and use this change as a stability metric to determine whether the local cognition matches the data distribution. When the metric shows that the stability of the learning process deviates from the preset range, the client will pause its action of uploading model update information to the server and actively request to resynchronize with the global model. This client-side autonomous early warning and refresh mechanism, combined with the server-side implicit semantic alignment and aggregation mechanism based on gradient projection, constitutes a federated learning system with a dynamic adjustment mechanism. This system can not only perceive the risk of cognitive failure of the local model without disclosing any specific data content that causes feature drift, but also prevent outdated local knowledge from affecting the integrity and timeliness of the global model through timely isolation and self-healing operations. Thus, it demonstrates continuous and reliable adaptability when dealing with long-term dynamic medical scenarios such as tumor evolution.

[0022] 3. By establishing a globally shared orthogonal semantic basis distributed by the server, this invention transforms the complex and costly explicit alignment negotiation process at the data or feature level in traditional federated learning into a gradient vector projection operation performed locally on the client with minimal computational cost, and a simplified aggregation and reconstruction operation of low-dimensional projection vectors performed on the server. This shift in information processing significantly reduces the amount of data that needs to be transmitted between the participants. Furthermore, since only the projection coordinates of the gradient on the public basis vector are transmitted, rather than the gradient itself, the method itself possesses a high level of privacy protection. Thus, this method avoids the high computational and communication costs associated with existing technologies that rely on complex cryptographic tools or multiple rounds of feature negotiation to achieve cross-domain collaboration. It provides a reliable technical path for deploying and running robust and efficient federated learning applications in large-scale medical institution networks with varying computing resources and technical capabilities. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0024] Figure 2 This is a performance comparison chart between the method of this invention and existing technologies in terms of recall rate for rare features;

[0025] Figure 3 This is a schematic diagram of the component deployment of the federated learning system of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] This invention provides a cross-domain federated machine learning training method for radiomics, applied to a federated learning system comprising a server and multiple clients. This system aims to construct a large-scale, cross-regional AI-assisted diagnostic model for medical imaging, such as a nationwide disease screening network covering multiple hospitals. The clients within the system are typically medical institutions in different regions, whose imaging equipment differs in manufacturer, model, and physical parameters, and whose local case data distribution also varies. The operational flow of this invention can be summarized into three mutually coordinating stages: local training and multidimensional information encoding on the client side, physical tracing on the server side... The clustering and semantic value-weighted aggregation stages, along with the final global model reconstruction and update distribution stages, enable the system to address the triple heterogeneity challenges stemming from device physical differences, uneven data distribution, and dynamic data drift. In a specific application scenario, when the federated task starts, to address the issue of numerous nodes in the network being unable to directly aggregate model updates due to feature space heterogeneity, the server first generates and distributes a set of globally shared gradient space implicit semantic coordinates during the federated learning system initialization stage. This aims to provide a unified and abstract semantic benchmark for all subsequent gradient information exchanges. Specifically, the server generates a set of [number missing] [unclear text - likely a number of coordinates]. a high-dimensional random vector, where The value of is a positive integer much smaller than the dimension of the model parameters, such as 16 or 32. To ensure that this set of basis vectors can form a semantic metric space without redundancy, the server then uses the Gram-Schmidt orthogonalization process to process this set of random vectors, thereby producing a set of globally orthogonal semantic basis vectors that are strictly orthogonal to each other. The set of basis vectors remains constant throughout the entire lifecycle of the federated learning task and is broadcast by the server to all participating clients at the start of the task. All clients then have a common reference frame, and the semantics of all subsequent local model updates will be measured and expressed by projecting onto this shared coordinate system.

[0028] Furthermore, during the operation of a federated network, a challenge arises from the inherent differences in the physical characteristics of the CT or MRI equipment used by different clients, particularly the point spread function (PSF). These differences can act as a systematic bias or device fingerprint, affecting all subsequently extracted radiomics features. To identify this physical bias before aggregation, this invention configures each client to actively detect and characterize the physical characteristics of the equipment before performing local training. The deterministic procedure is as follows: the client selects a raw medical image from its local dataset to be used for training. It then creates a digital copy of the image in memory and superimposes a standardized digital micro-perturbation signal (which can be a single-pixel pulse signal) onto a specific location of that copy, thereby generating a perturbation image. The client then sends the original image With disturbed images Simultaneously, the gradients of the feature maps generated by the two gradient maps are obtained, and the difference between them is calculated to obtain the vector magnitude of the gradient response difference. This scalar value is defined as a digital fingerprint that can characterize the physical characteristics of the client imaging device. Because the pattern and amplitude of this gradient response difference reflect how the local model responds to a standard pulse signal, and this response is modulated by the device's inherent PSF, therefore, This becomes an effective digital representation of the physical characteristics of a device, providing a basis for the server to group clients based on physical causes. At the same time, when dealing with long-term dynamic medical scenarios such as tumor evolution monitoring, the local data distribution of the client may experience feature drift. If not identified in time, the client may contribute outdated gradient information to the federated network.

[0029] This invention establishes a client-side autonomous early warning and refresh mechanism. Its operating principle lies in utilizing the dynamic changes in the loss function during local training. Specifically, during local model training, in addition to calculating the gradient used for backpropagation, the client also estimates the second derivative of the loss function with respect to the number of training steps, i.e., curvature. The curvature of the loss function reflects the learning stability of the model on the current data. Based on this, the client continuously calculates the variance of the loss function curvature within a sliding time window and uses it as a stability metric. The system presets a drift trigger threshold; when the client detects that the curvature variance exceeds this threshold... When the defined preset stability interval is reached, a local cognitive drift event is determined to have occurred. At this point, the client will suspend subsequent actions of uploading model updates to the server and instead send a resynchronization request to the server to obtain the latest global model, thus completing a local model refresh. Only when its stability metric is determined to be within the preset stability interval will the client proceed with the normal model update information submission process. After passing the stability check of local training, the client performs local model training and generates model update information. It performs one round of model training on its local heterogeneous data to obtain the original high-dimensional local gradient vector. To convey the semantics of model updates without transmitting the original gradients, the client performs a gradient-oriented projection operation, which specifically involves calculating the local gradient vector. Each basis vector in the globally orthogonal semantic basis previously distributed by the server The dot product is calculated, and the result is used as the projection length of the gradient onto the corresponding basis vector direction. In this way, the client can view the high-dimensional... Converted into a by A low-dimensional semantic projection vector consisting of projection lengths Ultimately, the client projects this low-dimensional semantic vector. Together with previously obtained digital fingerprints that characterize the physical properties of the device Upload them to the server together.

[0030] After receiving data packets from multiple clients, the server executes a two-level aggregation strategy. First, to address the systematic bias caused by device fingerprints, the server aggregates the digital fingerprints received from all clients. Clustering operations, such as using the K-Means clustering algorithm, are performed to group clients with similar imaging physical characteristics into one or more device clusters. Next, within each device cluster, the server independently performs weighted aggregation of the low-dimensional semantic projection vectors uploaded by clients within that cluster. This weighting aims to address the issue of unequal contribution values ​​from different clients in the network; the weights are based on each semantic projection vector. The concentration measure of the component distribution is calculated, specifically the information entropy. It is inversely proportional to the weight, information entropy The calculation formula is ,in, It is the number of basis vectors. yes The Middle The probability value obtained by normalizing the absolute value of each component, a client in a gradient semantic set, whose Sharp distribution, entropy value A lower gradient semantic projection value results in a higher weight, while a higher gradient semantic projection value results in a higher client entropy value and a lower weight. Through this weighted aggregation, the server generates an intra-cluster average semantic projection vector within each device cluster. Finally, the server fuses the intra-cluster average semantic projection vectors of each device cluster. For example, it can perform a weighted average of the intra-cluster average semantic projection vectors, where the weight is proportional to the number of clients in the cluster, thus obtaining a global fused average semantic projection vector. The server is based on this and its own globally orthogonal semantic base Through formula Reconstruct the global average gradient This calculation involves multiplying each component of the fused average semantic projection vector by its corresponding basis vector, and then summing all the resulting vectors; the server ultimately uses this reconstructed vector. This updates the global model and distributes the updated model parameters to all clients to begin the next round of federated learning training.

[0031] Example 1: In a federated learning network aimed at building a nationwide AI screening model for early lung cancer, client A is a large tertiary hospital located in a first-tier city equipped with a new model of Siemens CT scanner, and client B is a specialized research center located in a remote area using an older model of domestic CT scanner and possessing case data of a rare invasive lung cancer subtype. Under conventional federated learning aggregation methods, because client A has a much larger amount of data than client B, the gradient information contributed by client B for this rare disease subtype will be disproportionately diluted in semantics when weighted and averaged with the gradient of client A. This may result in insufficient sensitivity of the final global model for diagnosing this rare disease. In one cycle of the training method of this invention, client A and client B first obtain digital fingerprints that can characterize the physical characteristics of their respective imaging devices by superimposing standardized digital micro-perturbation signals onto digital copies of their local medical image data. and Due to differences in the physical imaging parameters of the two CT devices, their digital fingerprints... and Numerically, they exhibited distinguishable differences. Subsequently, both clients trained their models based on local data, obtaining their respective local gradient vectors, which were then projected onto the same set of globally orthogonal semantic bases distributed by the server to obtain low-dimensional semantic projection vectors. and Each element uploads its digital fingerprint and low-dimensional semantic projection vector to the server. Upon receiving the information, the server first performs a clustering operation based on the digital fingerprint. and Due to the differences, the server divides client A and client B into two different device clusters. This step physically isolates the gradient information originating from Siemens CT scanners and domestic CT scanners. Based on this, within their respective device clusters, the server performs weighted aggregation on the received projection vectors based on their information concentration. Client A, whose data content consists of common lung nodules, has a broader pathological semantic meaning in its gradients; therefore, its low-dimensional semantic projection vector... The component distribution is relatively uniform, and the calculated information entropy At a higher level, it correspondingly receives a lower weight when aggregating within a cluster; client B, on the other hand, has its data content highly concentrated on specific texture features of a rare and aggressive lung cancer subtype, and its gradient semantic orientation is clear, thus its low-dimensional semantic projection vector... The components exhibit a high concentration along a few basis vector directions, and the calculated information entropy... At a low numerical level, it obtains a higher weight when aggregating within a cluster. Through a two-level aggregation method that first groups based on physical causes and then weights the data within the cluster based on information value, the semantic direction of the rare pathological information contributed by client B is correspondingly preserved in the average semantic projection vector of its device cluster.

[0032] Ultimately, when the server fuses the average semantic projection vectors of each cluster and reconstructs the global gradient to update the model, this rare disease semantic information is integrated into the new global model parameters. Thus, the training method provided by this invention transforms the traditional problem of pursuing global feature space consistency into the problem of identifying and differentially integrating gradient contributions with different physical sources and different information values ​​within a shared abstract semantic space. This addresses the challenge of balancing the pursuit of model generalization ability with maintaining sensitivity to specific local features. After several rounds of updates, the global model generated by this federated learning network maintains the diagnostic ability for a large number of common cases represented by client A, while also possessing the ability to identify key features of rare invasive lung cancer subtypes contributed by client B. Moreover, this ability is obtained without directly accessing any client privacy data.

[0033] Example 2: To objectively verify the sensitivity preservation ability of the training method proposed in this invention for key rare pathological features under heterogeneous devices and imbalanced data environments, this example constructs a federated learning simulation test platform containing a server and 20 simulated clients. The 20 clients in the platform are divided into two device clusters, with 10 clients in each cluster, to simulate two CT devices with different physical imaging characteristics. This difference is simulated by applying a weak but different Gaussian blur kernel to each data to simulate the device-specific point spread function. The data source uses the publicly available LIDC-IDRI lung image dataset and performs imbalanced processing. In each device cluster, 8 clients are set as generalist nodes, whose local data consists almost entirely of common lung nodule images, while the remaining 2 clients are set as specialist nodes, whose local data consists of half common nodules and the other half of a rare malignant nodule subtype image enhanced by the model and accompanied by specific subtle texture features. A key parameter in the experiment is the number of globally orthogonal semantic basis vectors. The setting of this parameter requires a trade-off between the expressive power of the gradient semantic space and the communication overhead of the client uploading information; a parameter that is too small... A value may not adequately capture the complex pathological semantics, while an excessively large value... Values ​​that increase communication overhead and reduce training efficiency will increase the communication burden and reduce the efficiency of training parameters. The value of is related to the inherent complexity of the clinical problem to be solved. For the nodule subclassification task in this embodiment, after offline testing, the parameter ... The value of 16 to 32 can achieve stable convergence, therefore this experiment selects 16 as the optimal value. As a representative engineering value that balances expressive power and communication efficiency, the experiment included a control group. The control group was trained using the standard FedAvg algorithm, which aggregates gradients by weighting the original gradients based on the amount of local data on each client. The experimental group used the training method of this invention. Both groups were trained for 100 communication rounds under the same platform configuration and data distribution. After each round, an independent test set containing a balanced number of common and rare nodule samples was used to evaluate the rare feature recall rate of the global model. During the experiment, the server of the experimental group using the method of this invention was able to perform K-Means clustering on the digital fingerprints uploaded by each client, dividing the 20 clients into two device clusters consistent with the initial settings. The FedAvg method of the control group did not have this step. In the subsequent model aggregation stage, the server of the experimental group weighted the low-dimensional semantic projection vectors within each device cluster based on information entropy. The low-dimensional semantic projection vectors from specialist nodes, which contain rare nodule information, exhibited lower information entropy due to their high semantic concentration. Therefore, it was given a higher aggregation weight. The experiment recorded the recall rate of the global model of the two groups for rare features in different communication rounds. Table 1 shows the experimental data records of the key rounds.

[0034] Table 1: Comparison of rare feature recall rates for the two methods in different communication rounds.

[0035]

[0036] Table 1 shows that the recall rate of the model for rare features in the experimental group using the method of this invention increased significantly with the number of training rounds and remained at a high level. In contrast, the recall rate of the control group using the FedAvg method increased slowly and its final performance was limited. The data indicates that the gradient information of specialist nodes in the FedAvg method is overwhelmed by the gradient information of a large number of generalist nodes. The two-level aggregation method of this invention, by first isolating the physical differences of the equipment and then amplifying the contribution of high information value, enables the semantic information representing rare pathologies to be effectively accumulated and expressed in the global model. The results of this experiment confirm that, in the simulated cross-domain heterogeneous medical scenario, the cross-domain federated machine learning training method for radiomics of this invention can more effectively integrate the unbalanced but diagnostically valuable data distributed in the network compared with the standard federated averaging algorithm, thereby improving the sensitivity of the final global model to rare key clinical features.

[0037] Example 3: This example combines Figures 1 to 3 This section describes a cross-domain federated machine learning training method for image omics, as follows: Figure 1 As shown, the process begins with the server initialization phase, which generates and distributes a global orthogonal semantic base to establish a shared gradient space implicit semantic coordinate system. Subsequently, on the client side, a digital fingerprint representing the physical characteristics of the imaging device is generated through device physical characteristic detection and characterization steps. Then, a local gradient vector is obtained during local model training. After that, the local training stability verification phase is entered to determine whether the curvature variance of the loss function is within a preset stable range. If the result is no drift, the client will pause uploading and request synchronization of the global model to deal with cognitive drift. If the result is stable, gradient projection and information encoding are performed, projecting the local gradient vector onto the semantic base. A low-dimensional semantic projection vector is generated. Finally, the client performs an upload encoded information operation, uploading the digital fingerprint and the low-dimensional semantic projection vector together to the server. After receiving the information, the server initiates a two-level aggregation process. First, it performs physical source clustering based on the digital fingerprint uploaded by the client, dividing the client into device clusters. Then, it performs intra-cluster weighted aggregation within each device cluster, weighting the projection vectors according to information concentration. Subsequently, it obtains the average semantic projection vector of each cluster through cross-cluster fusion, and then performs global gradient reconstruction and update to update the global model. Finally, the server distributes the updated model to all clients to start the next round of training; and so on. Figure 2 As shown in the figure, the horizontal axis represents the number of communication rounds, and the vertical axis represents the rare feature recall rate. The solid line representing the method of this invention shows that its rare feature recall rate steadily increases with the increase of communication rounds, eventually reaching a level much higher than that of the dashed line representing the FedAvg method. This comparison objectively confirms the technical effectiveness of the method of this invention in integrating imbalanced data and maintaining sensitivity to rare key clinical features, as well as... Figure 3 As shown, the diagram contains a central server and multiple clients, such as the large hospital A and the specialist center B shown in the diagram. The central server is equipped with a global model, a global orthogonal semantic base, a client information database fingerprint and projection vector, and a core two-dimensional aggregation module. Each client is equipped with its own independent local model and local private image data, and is also equipped with a digital fingerprint generation module, a gradient projection module, and a stability monitoring module.

[0038] Example 4: In a federated learning network for monitoring the long-term evolution of chronic obstructive pulmonary disease (COPD), one of the client hospitals performed a routine software system upgrade on its main CT equipment. This upgrade introduced a new image reconstruction algorithm, which, without changing the core diagnostic features, produced a systematic and subtle change in the background noise distribution and micro-texture patterns of the images. This change constitutes a slow feature drift in the local data distribution of the client. If it cannot be identified in time, the local model will mislearn this equipment-related artifact as a pathological feature related to disease progression during subsequent training, thus contributing misleading gradient information to the federated network. To address this challenge, the method of this invention configures a stability metric and autonomous early warning mechanism based on the curvature change of the loss function on the client side. The key drift trigger threshold is not a fixed preset value, but is generated through an online calibration process. Specifically, in the first few communication rounds after the federated task is started, i.e., the first 20 communication rounds, the client system is configured in a baseline calibration mode.

[0039] In this mode, the client calculates and records the variance of its loss function curvature within a sliding time window during each local training iteration. After 20 communication rounds of calibration, the system uses all recorded variance values ​​to calculate its statistical mean. with standard deviation Therefore, the drift trigger threshold for this client is set to This threshold establishes a statistical upper bound on the normal fluctuation range of the client under stable training conditions. After the CT equipment software upgrade is completed, the client continues to participate in federated learning. When its local model begins to process image data generated by the new algorithm, the learning process of the model produces slight instability due to changes in data distribution. This manifests as an increase in the ruggedness of the loss function surface, which directly causes the variance of its loss function curvature to intermittently exceed the calibrated drift trigger threshold. When the client's monitoring module records that the number of times the curvature variance exceeds this threshold consecutively reaches a preset confirmation value, i.e., three consecutive batches, the system determines that a continuous local cognitive drift event has occurred. Once drift is determined to have occurred, the client will automatically pause uploading its local data to the server. The low-dimensional semantic projection vector obtained after gradient vector projection avoids the impact of updating information that may contain incorrect semantics on the global model. At the same time, the client sends a resynchronization request to the server to obtain the latest global model jointly maintained by other clients in the network that have not experienced data drift. After receiving the new global model, the client uses it as a starting point to retrain locally on the new data distribution where drift has occurred. Its stability measurement mechanism also enters a new round of baseline calibration to adapt to the new data state. This client-side drift warning and model refresh mechanism based on dynamic self-calibration threshold enables the entire federated learning system to have the ability to autonomously perceive and self-correct the long-term slow data dynamics of each node without the intervention of a central server.

[0040] Example 5: During the initial deployment of a federated learning network, the central server lacks prior knowledge of the types and quantities of imaging devices used by the numerous clients that will be connected. If a fixed cluster size K is preset for the K-Means clustering algorithm based on digital fingerprints for subsequent client clustering, this setting lacks objective basis and may lead to devices with different physical characteristics being incorrectly grouped into the same cluster or devices with similar physical characteristics being over-segmented, thus affecting the effectiveness of subsequent gradient aggregation. To address this challenge, this invention configures the server to perform a network device group discovery and clustering parameter adaptation before the federated learning task officially begins. The protocol should be defined in which the server first broadcasts an instruction to all registered clients, requiring each client to perform only the aforementioned steps of obtaining a digital fingerprint characterizing the physical properties of its imaging device and upload that digital fingerprint. After collecting the digital fingerprint set from all clients, the server uses a silhouette coefficient analysis method to determine the optimal number of clusters in the fingerprint dataset. Specifically, the server performs K-Means clustering on a series of K values ​​within a preset range, from 2 to an upper limit, and calculates the average silhouette coefficient corresponding to each K value. Finally, the K value that maximizes the average silhouette coefficient is determined by the system as the optimal number of clusters. This parameter is used as the final parameter for device clustering during subsequent federated training. This clustering parameter calibration process, which is performed before training based on the data distribution of the client group itself, enables the device clustering to accurately reflect the actual distribution of device types in the current network, avoiding the blindness of manually setting parameters. This calibration process can also be periodically retried by the server during the operation of the federated network when a large number of new clients join, thus providing a dynamic, adaptive, and reproducible operating basis for the two-level aggregation mechanism of the entire training method.

[0041] Example 6: Before deploying the federated learning training method of this invention in a newly established cross-regional medical collaboration network, a standardized pre-system parameter calibration procedure needs to be executed to address two specific engineering implementation problems. First, the intensity of the standardized digital micro-perturbation signal used to generate digital fingerprints needs to be determined to effectively stimulate measurable gradient response differences without significantly affecting the original image diagnostic information. Second, when performing weighted aggregation based on information entropy on the server side, the precise functional relationship between the information entropy value and the aggregation weight needs to be clarified. To calibrate the intensity of the digital micro-perturbation signal, the system uses an offline reference image dataset containing images of various typical lesions and normal tissues. The procedure uses a single-pixel pulse as the micro-perturbation signal and, within a preset intensity range (e.g., for 8-bit grayscale images, within the integer range of 1 to 255), progressively increases its intensity value and applies it to the reference image.

[0042] At each intensity level, the system calculates two indices in parallel: the vector magnitude of the gradient response difference between the perturbed image and the original image, and the structural similarity index (SSIM) of the perturbed image relative to the original image. Ultimately, the system selects the signal strength that allows the vector magnitude of the gradient response difference to reach 90% of its saturation value, while ensuring that the SSIM value is not lower than 0.99. This signal strength is then used as the globally unified and standardized digital micro-perturbed signal strength and is embedded in the system configuration files of all clients. To clarify the conversion relationship from information entropy to aggregation weights, the implementation of this invention concretizes the weight calculation rules in the weighted aggregation step into a specific mathematical expression. When the server performs weighted aggregation on the low-dimensional semantic projection vectors uploaded by each client within the device cluster, it first calculates the information entropy of each low-dimensional semantic projection vector according to the aforementioned method. Subsequently, the weight corresponding to this client in this aggregation. The result can be calculated using the following formula: ,in, It is a positive real number whose value is close to zero and is set to prevent division by zero errors, such as... This expression quantifies the inverse relationship between weight and information entropy without ambiguity. By calibrating the above two parameters and clarifying the calculation rules, the entire operation process of the federated learning training method achieves an engineering closed loop. All key control parameters and calculation logic that need to be pre-set in the system are defined deterministically, thus ensuring that the deployment of this technical solution in different scenarios has consistency and reproducibility.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cross-domain federated machine learning training method for radiomics, characterized in that, The method is applied to a federated learning system that includes a server and multiple clients, and the method includes: The server generates a set of globally orthogonal semantic basis vectors and distributes them to multiple clients to establish a shared gradient space implicit semantic coordinate system. Each client performs the following steps: Before training the local model, a standardized digital micro-perturbation signal is superimposed on a digital copy of the local medical image data to generate a perturbation image, and the gradients of the feature maps generated by the local medical image data and the perturbation image are extracted respectively, and then the difference in gradient response between the two is calculated to obtain a digital fingerprint. Models are trained based on their respective local data to obtain local gradient vectors. During local model training, a stability metric representing the curvature change of the loss function is obtained, and it is determined whether the stability metric is within a preset stable range. If the result is yes, the local gradient vector is projected onto each basis vector of the global orthogonal semantic basis to obtain a low-dimensional semantic projection vector. The digital fingerprint and the low-dimensional semantic projection vector are uploaded to the server, which performs the following steps: performing clustering operations on the digital fingerprints received from multiple clients to divide the clients into one or more device clusters; within each device cluster, weighted aggregation of the low-dimensional semantic projection vectors uploaded by the clients within that device cluster is performed to generate the intra-cluster average semantic projection vector of each device cluster, where each weight is calculated based on the concentration metric corresponding to the component distribution of the low-dimensional semantic projection vector. Finally, the server fuses the intra-cluster average semantic projection vectors and combines them with the global orthogonal semantic basis to reconstruct the global average gradient, which is used to update the global model.

2. The cross-domain federated machine learning training method for image omics according to claim 1, characterized in that, Concentration measure is information entropy Weights and Information Entropy Inversely proportional, information entropy Determined by the following formula: ,in, The number of globally orthogonal semantic basis vectors. For the client In the low-dimensional semantic projection vector, the first The probability value is obtained by normalizing the absolute value of each component.

3. The cross-domain federated machine learning training method for image omics according to claim 1, characterized in that, A standardized digital micro-perturbation signal is a single-pixel point pulse signal; a digital fingerprint is the vector magnitude of gradient response differences.

4. The cross-domain federated machine learning training method for image omics according to claim 1, characterized in that, The stability metric is the variance of the loss function curvature within a sliding time window; the preset stability interval is when the variance is less than the preset drift trigger threshold.

5. The cross-domain federated machine learning training method for image omics according to claim 1, characterized in that, When the stability metric exceeds the preset stability range, the client pauses uploading low-dimensional semantic projection vectors and sends a resynchronization request to the server to obtain the latest global model.

6. The cross-domain federated machine learning training method for image omics according to claim 1, characterized in that, The clustering operation performed by the server is the K-Means clustering algorithm.

7. A cross-domain federated machine learning training method for image omics according to claim 1, characterized in that, The client projects its local gradient vector onto each basis vector of the global orthogonal semantic basis by calculating the dot product of the local gradient vector with each basis vector to obtain the corresponding projection length as a component of the low-dimensional semantic projection vector.

8. The cross-domain federated machine learning training method for image omics according to claim 1, characterized in that, The server merges the average semantic projection vectors within each cluster in a step including: performing a weighted average of the average semantic projection vectors within each cluster to generate a merged average semantic projection vector, wherein the weights used for weighting the average are proportional to the number of clients contained in the cluster.

9. A cross-domain federated machine learning training method for image omics according to claim 1, characterized in that, Global orthogonal semantic basis vectors are generated by the server during the initialization phase of the federated learning system by performing a Gram-Schmidt orthogonalization process on a set of randomly generated vectors.

10. A cross-domain federated machine learning training method for image omics according to claim 1, characterized in that, The steps for the server to reconstruct the global average gradient are as follows: each component of the fused average semantic projection vector is multiplied by the basis vector of its corresponding global orthogonal semantic basis, and then all the resulting vectors are summed.

Images